Research map / Multi-agent systems and games / Human-AI and goal-conditioned
End-to-End Autonomous Driving Through V2X Cooperation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes UniV2X, an end-to-end vehicle-infrastructure cooperative driving framework that shares sparse agent/lane queries and a dense occupancy probability map to improve perception, mapping, occupancy, and planning on DAIR-V2X.
Problem
Cooperative (V2X) autonomous driving research optimizes individual modules rather than final planning; end-to-end VICAD must handle limited bandwidth, latency-induced misalignment, and untrusted transmitted data.
Approach
A query-based unified network that transmits sparse instance-level queries (agents, lanes) and a dense but low-channel occupancy probability map (with flow prediction for future steps), with rotation-aware query transformation, Hungarian-based query matching, and occupancy max fusion, so that all modules and planning are jointly optimized.
Key results
- UniV2X reduces average collision rate from 0.89% (No Fusion) to 0.34% on DAIR-V2X while cutting transmission cost to 8.09e5 BPS versus 8.19e7 for Vanilla.
- Detection mAP rises from 0.165 (No Fusion) to 0.295 and AMOTA from 0.163 to 0.239 for the car class.
- Online mapping IoU-Lane improves from 6.4% to 17.8% and IoU-Crosswalk from 2.7% to 19.8%; occupancy IoU-f improves by +13.4 points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UniV2X instead of No Fusion | Cooperative autonomous driving · DAIR-V2X | average collision rate (%): 0.89 -> 0.34 | Helped |
| Use UniV2X instead of BEV Feature Fusion | Cooperative autonomous driving · DAIR-V2X | average collision rate (%): 0.83 -> 0.34 | Helped |
| Use UniV2X instead of No Fusion | Agent perception · DAIR-V2X (car) | mAP: 0.165 -> 0.295 | Helped |
| Use UniV2X instead of V2X-ViT | Agent perception · DAIR-V2X (car) | AMOTA: 0.287 -> 0.239 | Hurt |
| Use UniV2X instead of No Fusion | Online mapping · DAIR-V2X | IoU-Lane (%): 6.4 -> 17.8 | Helped |
| Use UniV2X instead of No Fusion | Occupancy prediction · DAIR-V2X near region | IoU-n (%): 16.3 -> 22.0 | Helped |
| Add agent fusion (ablation) | UniV2X · DAIR-V2X agent perception | mAP (%): 16.5 -> 29.5 | Helped |
| Add lane fusion (ablation) | UniV2X · DAIR-V2X online mapping | IoU-Lane (%): 6.4 -> 17.8 | Helped |
| Add occupancy fusion (ablation) | UniV2X · DAIR-V2X occupancy prediction | IoU-f (%): 13.1 -> 23.5 | Helped |
| Add flow prediction rather than removing the prediction module (ablation) | UniV2X · DAIR-V2X agent perception, 500 ms latency | mAP (%): 25.9 -> 28.5 | Helped |
| Increase communication bandwidth from 0.3 Mb/s to 1.0 Mb/s (ablation) | UniV2X bandwidth-constrained · DAIR-V2X | mAP (%): 21.0 -> 29.5 | Helped |
| Use UniV2X instead of No Fusion | Cooperative autonomous driving · V2X-Sim | average collision rate (%): 2.0 -> 1.75 | Helped |
Surprising findings
- CooperNaut has a lower off-road rate but a much larger L2 error (5.35 m average), attributed to a conservative short planning horizon.
- V2X-ViT achieves higher AMOTA (0.287 vs 0.239) than UniV2X but at far higher transmission cost, and compressing it to similar cost drops AMOTA to 0.071.
Limitations
- Only open-loop evaluation; closed-loop experiments left for future work.
- Several interaction fusion modules remain preliminary, and the framework is complex.
- Ego-vehicle position errors can corrupt the ego-area removal step; training with inference-style queries is still being explored.
Experimental setup
Datasets: DAIR-V2X V2X-Sim nuScenes
Benchmarks: DAIR-V2X V2X-Sim
Metrics: L2 error collision rate off-road rate BPS transmission cost mAP AMOTA IoU-lane IoU-crosswalk IoU-n IoU-f
Baselines: No Fusion Vanilla BEV Feature Fusion CooperNaut Early Fusion Late Fusion V2X-ViT Where2comm DiscoNet CoAlign CoCa3D CoBEVT
Builds on: UniAD BEVFormer DETR Panoptic SegFormer Where2comm
Closest papers by meaning
- Privacy-Preserving V2X Collaborative Perception Integrating Unknown Collaborators (AAAI 2025 · similarity 0.87)
- Bench2Drive: Towards Multi-Ability Benchmarking of Closed-Loop End-To-End Autonomous Driving (NeurIPS 2024 · similarity 0.87)
- DriveTransformer: Unified Transformer for Scalable End-to-End Autonomous Driving (ICLR 2025 · similarity 0.87)
- MAPLE: Latent Multi-Agent Play for End-to-End Autonomous Driving (NeurIPS 2026 · similarity 0.87)
- Learning Cooperative Trajectory Representations for Motion Forecasting (NeurIPS 2024 · similarity 0.86)
- Autonomous Driving Research Requires a Community-Driven Data Paradigm (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.